Making wearable activity recognition robust & lightweight with a two-stage reservoir computing model

What is it about?

Wearable motion sensors (IMUs) can help detect what activity someone is doing, but models often break when the device is worn differently or when the user is a new person the model hasn’t seen before. In the 2nd WEAR Dataset Challenge, we had to recognize 18 activities (plus a dominant “null/rest” class) from 1-second sensor windows captured at four body locations, with data augmentation that mimics real-world wearing conditions. We propose a two-stage hybrid pipeline. First, a binary CNN identifies whether a window is “null” or “activity.” Only windows predicted as activity go to the second stage, where a multiclass CNN predicts the specific activity. To represent each 1-second window well, we combine: 1) statistical features from time, frequency, and wavelet domains 2) reservoir computing embeddings that capture short-term motion dynamics. The reservoir part uses multiple small reservoirs (echo-state style) applied across IMU axes to produce high-dimensional temporal features that complement standard statistics.

Featured Image

Photo by i yunmai on Unsplash

Photo by i yunmai on Unsplash

Why is it important?

In real-world wearable sensing, models must handle sensor misalignment, noise, class imbalance, and differences between people. The WEAR challenge makes this especially hard by using very short (1-second) windows and evaluating on participants not seen during training, which limits how much temporal context a model can use. Our results show that two design choices matter a lot under these conditions: 1) Two-stage “null filtering” helps explicitly removing “null/rest” first reduces confusion and improves downstream recognition. 2) Reservoir computing features add useful time dynamics (i.e. reservoir embeddings) which help capture short-term motion patterns in 1-second windows, complementing statistical features and improving robustness.

Perspectives